A spatiotemporal attention autoencoder (STAAE) for early-stage corrosion detection in carbon steel using eddy current pulsed thermography

Detection of early corrosion in carbon steel using Eddy Current Pulse Thermography (ECPT) faces the challenges of weak signal and noise interference. This paper proposes a spatiotemporal attention autoencoder (STAAE) that integrates advanced temporal and spatial feature extraction techniques. Sequential data is processed through an LSTM layer and spatial information is extracted through a convolutional pooling layer and combined with a self-attention mechanism that dynamically adapts to critical temporal variations. The model is trained on non-corrosive data to learn feature patterns of normal ECPT signals for subsequent corrosion detection. The model generates reconstruction errors when corroded regions are tested to localise corroded areas reliably. This paper shows the effectiveness of the proposed method in visualising early corroded areas.